Interactive activity recognition using articulated-pose features on spatio-temporal relation

Abstract

A success progress of pose estimation approaches motivates the activity recognition used in CCTV-based surveillance systems. In this paper, a method is proposed for recognizing interactive activities between two human objects. Based on articulated joint coordinates obtained from a pose estimation algorithm, the distance and direction feature are extracted from objects to describe both the spatial and temporal relation. The multiclass Support Vector Machine is finally employed for activity classification task. Compared with existing methods using the public interaction dataset, the proposed method outperforms in overall classification accuracy.